There has been a lot of hullabaloo lately around the ‘collapse of the GenAI bubble’, primarily driven by the plummeting valuations of its bellwethers – Nvidia, Microsoft and other Big Techs riding atop the AI tiger. Intel stock plunged a vertiginous 25%, Microsoft and Google both dipped on their seemingly stellar earnings reports, and even the mighty Nvidia plunged hundreds of billions of dollars from its dizzying high of $3.2trn valuation. Investors got unnerved when these companies like Microsoft and Google committed to spending upwards of $50bn a year each every year, with a disproportionate amount going to data centres and cloud infrastructure underpinning AI. The fact that they could not show any immediate outlook of returns spooked the investor community, leading to a global mini-crash in markets worldwide.

Beyond the investor community, it led people to ask as to who was really making money from AI, if at all? In this column, let me dive into the same, using a favourite software engineering term here: lets us look at the Generative AI ‘stack’- the layers that make up the technology. The lowest layer of the stack (call it Layer 1) are the infrastructure, or infra, providers – companies like Nvidia and AMD which make the powerful GPUs which make GenAI Large Language Models (LLMs) work, companies like TSMC which build the nanometric chips that make them, and ASML which make space-age machines that make the chips . This layer, the picks of shovel providers of the gold rush, have made the most amount of money with Nvidia earnings increasing by eyepopping numbers every quarter, and TSMC and ASML assuming geopolitical significance along with making gobs of money.

Layer 2 above it is the Model layer, the one grabbing all the eyeballs. These are the OpenAIs, Anthropics and even the Googles and Metas of the world making the LLMs that seem to do magical things for us- ChatGPT, Claude, Llama, Gemini, etc. If the infra layer spouts out money, the model layer consumes all and most of it. The stratospheric costs of compute (the infra), data and talent is what compels the BigTechs to spend the billions, and smaller startups like OpenAI to ally themselves with the Microsofts of the world so that they can get the moneys required to build the models. This layer is the one worrying the investors, but is what is arguably building the future. Layer 3 above are the dev platforms, which are used by developers to build applications on top of the LLMs. Usually nestled in large clouds like Azure and AWS, this is the investment that the BigTechs do to get applications built on top of them.

The top layer 4 is what is called the application layer. This is where innovation abounds, with startups and big companies building their own focused GenAI apps on top of the powerful LLMs that Meta or OpenAI has constructed. These are vertical apps like the GenAI search app Perplexity, Copy.ai for marketers, Gamma.ai to make amazing presentations or Harvey which is a GenAI app for lawyers. The money required is comparatively trifling, leading to hundreds of innovative apps being built every year. But there is no moat, with the IP mostly belonging to the LLM providers, and the continual threat of the LLM creating a similar application. The only competitive advantage here is how well you go-to-market, and how good your interface is. This layer is the future of GenAI, but currently makes and consumes little money.

Most GenAI stacks also have the Cloud layer, which runs across the stack. The clouds are owned by the hyperscalars – AWS by Amazon, Azure by Microsoft, and Google Cloud – which subsume many of the above layers. The GPUs, LLMs, dev platforms and applications both power and run on this ubiquitous cloud, and this is what is responsible for the lip-smacking earnings of the BigTech players. These are incestuous relationships – Microsoft has invested $13bn in OpenAI, for instance, but all GPTs run exclusively on the Azure cloud, so Microsoft makes money irrespective of whether a customer buys GenAI from OpenAI or Microsoft.

There are two layers, however, which no stack I have seen talk about. There is a layer below the infra layer, call it Layer 0, and this is the real infrastructure that powers GenAI. This is the electricity generators, the water providers, the massive data centre constructors. For example, the electricity needed to power the data centre behemoths will be more than what a Germany or Japan consumes very soon, and the large utilities that make up this layer are raking it in too. Finally there is a layer right at the very top, Layer 5, are the consulting and knowledge firms which have jumped on the GenAI bandwagon to buff up their own earnings. BCG estimates 20% of its revenues to come from GenAI related work, that is approx.. $2.5bn, and Accenture claims to have booked $2bn GenAI revenue. As large corporations retool and redesign their organisations and employees this layer will continue to grow.

Whether GenAI is making or not, therefore, is not a simple yes/no answer, it depends on where do you sit in the value chain – a complex, ‘layered’ answer if it may.